Research On Semantic Segmentation Algorithm Of Urban Street View Image Based On Convolutional Neural Network

At present,there are two main ideas for the application of image semantic segmentation in the field of unmanned driving: One way of thinking is to focus on the model design of precision index,which leads to a huge amount of computation and a long time for segmentation prediction,which is difficult to deploy and implement in the practical application with limited resources;In the design of one kind of thinking,the main focus is on the reasoning speed of the model,but due to the loss of the underlying details and the reduction of the number of channels,the accuracy of a certain degree decreases.Now in the autonomous driving scenario,semantic segmentation task is moving towards the pursuit of an effective balance between the two,rather than just the pursuit of a certain aspect of accuracy or speed.This paper studies the above two directions respectively,and the main research contents and innovations are as follows:First of all,in order to be closer to the real scene,urban street view pictures need high resolution,high quality clear pictures.Considering the limitation of computing resources and graphics card resources,In this paper,aiming at the problems of large amount of calculation and large number of parameters in existing high-precision models,Firstly,the backbone feature extraction network is improved and replaced on Deep Lab V3+ network;then,a lightweight semantic segmentation method integrating multi-scale features is designed to enhance the feature representation capability of a single convolution;aiming at the simple Deep Lab V3+ network encoder module,some important pixels are lost and pixels in the image are discontinuous after upsampling,which affects the accuracy of the network model,an effective multi-scale feature fusion module is designed.In the Cityscapes data set,the network reached 74.85% MIoU with only 20.52 Params.In the next place,driverless cars need to sense the surrounding environment including other vehicles,pedestrians,road signs,traffic lights and road conditions in real time,so as to make quick and accurate decisions to ensure safe and efficient driving.In this paper,we propose a lightweight real-time semantic segmentation network FDSANet,which integrates deeply separable convolution and attention mechanism,and is committed to seeking the balance between segmentation accuracy and running time of network.The RDS＿Stem Block with depth separable convolution is proposed as the feature extractor module in the first downsampling stage of DSANet feature extraction network;refer to the STDC module as the subsequent module part of the backbone feature extraction network in the encoder;S＿ARM module is designed to extract more rich high-level semantic features.Comparison experiments on urban Street view data set show that FDSANet network can achieve 71.85% MIoU with 97.23 FPS.